Skip to content
The Chat Archive
command-a-03-2025· Cohere LMArena Human Preference 140K

translate it to english, to a recruiter in a profesional tone """ I remember when I was…

  1. User

    translate it to english, to a recruiter in a profesional tone """ I remember when I was developing a recommendation system to recommend job offers to job seeker users (the parts) so in the very begining versions of the system we just started with the knowledge filters using the known features of both parts and computing the semantic similarity matches between the average embedding of the parts' hard skills, in this point the results were no as good as we expect, but we keep iterating and researching the process and we could use better embeddings models like the OpenAI's ADA and then the 3-large, getting better matches; then we increase the content filter computing the semantic similarity between the Job title and the user's preferred role name, and building a weighted avarage score using the two similarities, and also we could improve the knowledge filter extracting more and in a better way the job offer's features since we change the use of Name Entity Recognition of NLP technic to use LLM request """

  2. Assistant

    Here is the translated text in a professional tone suitable for a recruiter:


    During my work on developing a recommendation system designed to match job offers with job seeker profiles, the initial versions of the system relied on knowledge filters. These filters utilized known features of both the job offers and the candidates, calculating semantic similarity matches between the average embeddings of their hard skills. While the initial results did not meet our expectations, we continued to iterate and refine the process. By incorporating more advanced embedding models, such as OpenAI's ADA and subsequently the 3-large model, we significantly improved the quality of matches.

    To further enhance the system, we introduced a content filter that computed semantic similarity between the job title and the candidate’s preferred role. We then developed a weighted average score by combining the two similarity metrics. Additionally, we optimized the knowledge filter by improving the extraction of job offer features. This was achieved by transitioning from traditional Named Entity Recognition (NER) techniques in NLP to leveraging Large Language Model (LLM) requests, which yielded more accurate and comprehensive results.


    This version maintains a professional tone while clearly conveying the technical improvements and iterative process of the project.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.